Partial Clone for Large Repositories

CAUTION: Alpha: Partial Clone is an experimental feature, and will significantly increase Gitaly resource utilization when performing a partial clone, and decrease performance of subsequent fetch operations.

As Git repositories become very large, usability decreases as performance decreases. One major challenge is cloning the repository, because Git will download the entire repository including every commit and every version of every object. This can be slow to transfer, and require large amounts of disk space.

Historically, performing a shallow clone (--depth) has been the only way to reduce the amount of data transferred when cloning a Git repository. This does not, however, allow filtering by sub-tree which is important for monolithic repositories containing many projects, or by object size preventing unnecessary large objects being downloaded.

Partial clone is a performance optimization that "allows Git to function without having a complete copy of the repository. The goal of this work is to allow Git better handle extremely large repositories."

Specifically, using partial clone, it should be possible for Git to natively support:

large objects, instead of using Git LFS
enormous repositories

Briefly, partial clone works by:

excluding objects from being transferred when cloning or fetching a repository using a new --filter flag
downloading missing objects on demand

Follow Git for enormous repositories for roadmap and updates.

Enabling partial clone

GitLab 12.1 uses Git 2.21.0 which has an arbitrary file access security vulnerability when uploadpack.allowFilter is enabled, and should not be enabled in production environments.

A feature flag is planned to enable uploadpack.allowFilter and uploadpack.allowAnySHA1InWant once the version of Git used by GitLab has been updated to Git 2.22.0.

Follow this issue for updated.

Excluding objects by size

Partial Clone allows large objects to be stored directly in the Git repository, and be excluded from clones as desired by the user. This eliminates the error prone process of deciding which objects should be stored in LFS or not. Using partial clone, all files – large or small – may be treated the same.

With the uploadpack.allowFilter and uploadpack.allowAnySHA1InWant options enabled on the Git server:

# clone the repo, excluding blobs larger than 1 megabyte
git clone --filter=blob:limit=1m <url>

# in the checkout step of the clone, and any subsequent operations
# any blobs that are needed will be downloaded on demand
git checkout feature-branch

Excluding objects by path

Partial Clone allows clones to be filtered by path using a format similar to a .gitignore file stored inside the repository.

With the uploadpack.allowFilter and uploadpack.allowAnySHA1InWant options enabled on the Git server:

Create a filter spec. For example, consider a monolithic repository with many applications, each in a different subdirectory in the root. Create a file shiny-app/.filterspec using the GitLab web interface:

# Only the paths listed in the file will be downloaded when performing a
# partial clone using `--filter=sparse:oid=shiny-app/.gitfilterspec`

# Explicitly include filterspec needed to configure sparse checkout with
# git config --local core.sparsecheckout true
# git show master:snazzy-app/.gitfilterspec >> .git/info/sparse-checkout
shiny-app/.gitfilterspec

# Shiny App
shiny-app/

# Dependencies
shimmery-app/
shared-component-a/
shared-component-b/

Create a new Git repository and fetch. Support for --filter=sparse:oid using the clone command is incomplete, so we will emulate the clone command by hand, using git init and git fetch. Follow gitaly#1769 for updates.

# Create a new directory for the Git repository
mkdir jumbo-repo && cd jumbo-repo

# Initialize a new Git repository
git init

# Add the remote
git remote add origin git@gitlab.com/example/jumbo-repo

# Enable partial clone support for the remote
git config --local extensions.partialClone origin

# Fetch the filtered set of objects using the filterspec stored on the
# server. WARNING: this step is slow!
git fetch --filter=sparse:oid=master:shiny-app/.gitfilterspec origin

# Optional: observe there are missing objects that we have not fetched
git rev-list --all --quiet --objects --missing=print | wc -l

CAUTION: IDE and Shell integrations: Git integrations with bash, zsh, etc and editors that automatically show Git status information often run git fetch which will fetch the entire repository. You many need to disable or reconfigure these integrations.

Sparse checkout must be enabled and configured to prevent objects from other paths being downloaded automatically when checking out branches. Follow gitaly#1765 for updates.

# Enable sparse checkout
git config --local core.sparsecheckout true

# Configure sparse checkout
git show master:snazzy-app/.gitfilterspec >> .git/info/sparse-checkout

# Checkout master
git checkout master